A new study has shown that agentic AI browsers, which can independently shop, book tickets, and manage accounts, are highly vulnerable to online threats. Researchers at Guardio tested Perplexity’s Comet, the most prominent agentic AI browser, and found it could be tricked into buying fake products, falling for phishing scams, and downloading malicious files. In one case, Comet completed a purchase on a fake Walmart site without verifying its authenticity or asking for user approval.
Another test revealed that it treated a phishing email as legitimate, loading a fake Wells Fargo page and requesting login details. A third test used hidden instructions on a mock CAPTCHA page to force Comet into downloading malware. These flaws highlight how attackers could exploit a single AI model repeatedly, scaling attacks with little effort. Experts caution that agentic AI browsers are not yet secure enough for sensitive activities like banking or shopping and advise users to avoid providing personal or financial details to them.
Read more...